Page 55 - IJEEE-2023-Vol19-ISSUE-1
P. 55
Al-Qaysi1, Al-Saegh, Hussein & Ahmed | 51
whereas ,(-) stands for the time-series signal, a stands for input samples to train big-scale networks. Therefore, transfer
dilation, and b represents the translation factor. The .!,#(t) learning is adopted to tackle that problem [33]. Transfer
represents the complex conjugate and can be calculated by learning means the use of an already-trained network in
solving another classification problem by re-train a few
.!,# (-) = 1 . 3- - '5 (2) numbers of its last layers. This saves a lot of time for training
1|%| % and requires fewer training samples than training the network
from scratch. In this regard, VGG-16 [33] is a famous CNN
where ?(t) is the wavelet. The major weakness of CWT is model with 16 convolution layers proposed by Oxford Visual
that both dilation and translation parameters change Geometry Group in 2014 and it has achieved a outstanding
continuously. Hence, the wavelet’s coefficients for all performance in numerous image processing tasks. In this
available scales after calculation require numerous efforts study, the VGG-16 model is used for the MI EEG
but yields inconsequential information [32]. The wavelet classification problem. The generated scalograms are used
transform method can be considered as a mathematical for training the network.
microscope that splits up a signal into a bunch of signals. In
such a method, the same signal corresponding to different E. WTNN Evaluation Metrics
frequency bands can be represented to provide frequency
bands at appropriate time intervals. Three types of mother The performance of the proposed WTNN system has
wavelets namely Morlet, Bump, and Amor were used in this been evaluated using seven metrics namely accuracy,
study precision, sensitivity, specificity, F1 score, LogLoss, and
AUC. TABLE I gives the mathematical equations for each
D.Deep Feature Extraction and Classification of the metrics with a brief description of each of them [34].
In the table TPL: true positive, TN: true negative, FP: false
The classification problem of EEG signals requires high positive, and TN: true negative.
dimensional features for representing the latent features of
the brain signal. Since the classification method plays a The Receiver Operating Characteristics (ROC) curve is
major role and has a direct impact on the discrimination also used for measuring the performance of the models. The
between two MI EEG mental commands, therefore, the curve is used for checking the performance of the
selection of an appropriate classifier is crucial. The classical classification model at various threshold settings by
machine learning methods need hand-crafted features to distinguishing between classes (i.e., a degree of
perform classification. Deep CNN (DCNN), on the other separability).
hand, performs classification by extracting features directly
from the raw data [33]. To evaluate how well a model will perform on unseen
MI EEG inputs, k-fold cross-validation was used in this
CNN depends on the convolution process in extracting study. In k-fold cross-validation, the data is divided into k
dominant features by adopting several kernels (also known subsets, in which k-1 subsets are used for training the model
as filters). Small kernels are moved horizontally and and the residual subset is used for testing the model. This
vertically along the input sample to capture important process is repeated for k times (folds) until all the subsets are
features which will be translated as coefficients in those used as validation data. The results obtained from the k-folds
kernels. However, deep learning requires a lot of time and can be averaged to determine the accuracy of estimation.
This study used the 10-fold cross validation for the training
TABLE I
THE USED EVALUATION METRICS
Evaluation metric Mathematical equation Explanation
#7 + #9
Classification accuracy !6 = #7 + :7 + :9 + #9 The ratio of the number of correctly classified samples to the
total number of the same class input samples
Precision Precision = #7 The number of correctly classified samples among all the
#7 + :7 classified samples. It tests the classifier's ability to reject
irrelevant subjects.
#7
Recall (Sensitivity) Recall = #7 + :9 The number of correctly classified samples from all the
positive representations.
F1-score :1?score = 2 * 2* #7 + :9 The F1 score can be described as a weighted average of
#7 + :7
precision and recall, where an F1 score achieves its best
#9 value at 1 and the worst value at 0.
#9 + :7
Specificity Specificity = Assesses a model’s ability to detect true negatives of each
Log Loss category.
(
LogLoss = - 1 [B) log, (BF) ) Log loss is the crucial classification metric based on
> ? probabilities. It defines the probability outputs of a classifier
instead of its discrete predictions.
)*+
+ (1 - B))log,(1 - BF))]
n: number of samples
yF-: predicted probability per label